Summary

The first business case from Career Sharing Day was that carving out a narrow niche within a broad track is the easiest way to build distinctiveness. “Military from a God’s-Eye View” had no military experience, but analyzed military bases using open-source satellite maps such as Google Earth, moving from CEO of an advertising agency to independent creator with roughly one million followers across Douyin, Bilibili and YouTube. The transition was painful; he also said self-media requires social insight, and that monetization includes advertising, without specifying the ad formats.
The gaming example framed the limits of AI coding in terms of system complexity: generating a simple demo does not mean delivering an open world. Word-matching monster-fighting games and Snake-style games should already be easy to generate, but the speaker thought Minecraft was probably beyond what AI could write; lighting and reflections still require handling angles, weather, time of day and material roughness. Another Assassin’s Creed bug involved low frame rates, collisions and inaccessible boundaries. “The more open it is, the harder it is to build, because once you stop constraining it, it gets hard.”
庄明浩’s core teaching was not to mythologize AI, but to define it as “a super-parrot that has read every book in the world.” Data, algorithms and compute produce a Sun Wukong that appears able to do anything from heaven to earth, but underneath, it is still predicting the most likely next character and does not truly understand that one plus one equals two. That is why it produces hallucinations, bias and common-sense errors, and “has no feelings and does not understand right from wrong.”
The minimum operating system for output quality is “role, task, requirements,” and the easiest element to overlook is requirements. The same explanation of a black hole can produce a pile of technical jargon if the prompt simply says “explain black holes”; ask an astronomer to explain it to a sixth grader in plain language and with analogies, and the result changes completely. Format, word count, grade level, tone and whether to use analogies all narrow the answer distribution.
The real red line in learning is “asking directly for the answer,” because that outsources thinking itself. A better use is to have AI act like a teacher by asking questions step by step, explaining unfamiliar concepts, correcting grammar and offering essay angles, without writing the body or supplying the final answer. “It will keep getting stronger, but you will not keep getting stronger; you may actually get dumber” is 庄明浩’s injunction to students.
AI has started replacing parts of execution, rather than waiting to eliminate an entire profession in one shot. Drawing a winged refrigerator by hand might once have taken a professional 3 days with repeated revisions; now it may take seconds. NotebookLM has also gradually taken over part of 庄明浩’s PPT production. But after comparing this with the camera’s displacement of some portrait-painting demand, he acknowledged plainly: “There is no answer to this question.”
Doubao Phone demonstrated the shift from a chat box to direct task execution, provided it has app permissions and can operate visibly or in the background. The live instruction asked it to open Xiaoyuzhou, find the first episode of “No Kidding” and like it; faced with 200-300 episodes, it advanced about 3 episodes per screen and eventually completed the task on its own. More app permissions could expand its range, but the demo’s execution speed was still very slow, and users can choose whether to watch it operate.

Deep dive

1. Career Sharing Day Shifted from Showcasing Jobs to Offering an Understandable Entry Point

小庄 attended 2 career-sharing sessions for fifth-grade parents: one on military self-media and one on games. In previous years, each grade typically had only 5 or 6 options; this time, fifth and sixth grades were scheduled separately, yet the venue still displayed a large number of roll-up banners, with possibly more than 10 topics in total.
After discussing it with the homeroom teacher, 庄明浩 decided not to give a broad overview of AI, but to narrow the topic to “how to use AI as a learning tool.” The under-40-minute session retained only 2 blocks: why AI looks so powerful, and how students can use it well. He then compressed the message into 2 takeaways: “You are Tang Sanzang; AI is Sun Wukong,” and “role, task, requirements.”

2. Military Self-Media Built Content Distinctiveness Through a Narrow Niche

The creator 小庄 introduced is called “Military from a God’s-Eye View.” He never served in the military, but uses open-source satellite maps such as Google Earth to analyze military affairs and military bases. He was previously CEO of an advertising agency, then went through a “very painful” career transition; today, he runs the self-media operation himself.
The creator’s method for choosing topics is to identify a broad track first, then find a sufficiently narrow niche within it. He does not discuss military affairs broadly, but focuses exclusively on “open-source satellite military analysis.” 庄明浩 immediately pressed on the commercial result: with roughly one million followers across Douyin, Bilibili and YouTube, “he should be making pretty decent money.”
Monetization includes advertising, but the specific ad formats were not explained. Discussing advertising, the creator said the work requires creativity; discussing self-media, he listed social insight as one of its basic qualities. He also noted that when large-screen phones and mobile advertising were not yet widespread, newspaper ads were an important way for consumers to learn about products and develop interest in them.

3. Game Education Showed the Finished Product Without Fully Explaining the Production System

The second speaker, addressing students with no exposure to game development, used a simple word-matching monster-fighting game to demonstrate a code panel, and also reviewed the evolution of games such as Snake and Mario. 小庄 called the PPT “not very deep,” prompting 庄明浩 to joke: “I’m no longer a professional, but why is he even less professional than I am?”
The difference between the 2 game presentations was clearly stated: this one was more about explaining “what games look like behind the scenes,” while 庄明浩’s presentation the previous year was about “how to make games”—what the roles are, how the workflow connects, what each type of employee does, and how the pieces ultimately become a complete deliverable.
The speaker acknowledged that AI can write code, saying that when he tried making a game at the time, AI “only wrote the final 1,000 lines of code.” But he thought Minecraft was probably beyond what AI could write because the code was simply too long. 庄明浩, meanwhile, thought the word-matching demo shown in the room should already be very easy to produce.

4. The Real Cost of Games Lies in Lighting, Collisions and Boundary Conditions

The speaker used a long formula filled with Greek letters to break down lighting and reflections: the angle of indoor lighting, whether an outdoor scene takes place in the morning, noon or afternoon, rain and overcast conditions, and the roughness of an object’s surface can all change the final image. Students may not have understood the formula, but they could see the computational load behind a polished visual.
He also analyzed an Assassin’s Creed bug he had encountered. He speculated that there might have been an impassable wall; the carriage may have been moving too fast while the game’s frame rate was too low, so in a later frame it had already passed through the wall. The system may then have pulled the carriage back, and after pulling too hard, sent it crashing into a character. This was his speculation, not a confirmed conclusion from the source code.
庄明浩 used this to explain the “invisible walls” in games such as Black Myth: Wukong: players can see an area, but cannot enter it because the content beyond may be unfinished, or entering it may trigger more complex errors. Boundaries are an important design mechanism for keeping players inside the intended range and avoiding potential failures.
Minecraft is not a fixed map. It uses a seed to continuously generate new areas outward from a local region, with terrain produced randomly, making it difficult for players to move faster than the map can generate. 庄明浩 called this “a kind of shortcut, to some extent,” and summarized it as follows: “The more open it is, the harder it is to build, because once you stop constraining it, it gets hard.”

5. Explaining Things to Children Starts with Finding Shared Cultural Ground

庄明浩 initially asked AI to create teaching materials themed around Naruto, assuming students would be familiar with it. 小庄 pointed out that many students might only have heard the name without knowing the story. A poll in the room confirmed that many “didn’t really know it,” so the theme was changed.
The second version switched to Eggy Party, which more people were likely to know, but some students still did not. The third version ultimately chose Journey to the West, because 庄明浩 thought it was unlikely that anyone would not know what it was about. The 3 revisions showed that analogies must be built on knowledge the audience already has.
The PPT itself was produced with help from NotebookLM. 庄明浩 called it Google’s latest software, designed specifically for making images and PPTs, and said it was already “very good.” For someone who produces large volumes of presentation files every day, it is gradually replacing part of the PPT production work, but topic selection and audience calibration still require a human.

6. November 30, 2022 Became a Cutoff Point in the Generational Debate over AI

庄明浩 first placed AI on a timeline spanning decades: researchers had begun studying related questions around the 1940s and 1950s. Siri, smart speakers, in-car voice assistants and robot vacuums are all examples of the previous generation’s capabilities entering daily life, though their conversations were often mechanical and prone to getting stuck.
A robot vacuum can now execute “don’t clean the bedroom; only clean the living room and kitchen,” provided it has identified the rooms and built a household map. 庄明浩 used this to show that the current generation of AI can understand environments and execute more specific, constrained tasks—something robot vacuums from a few years ago could not do.
He used ChatGPT’s launch on November 30, 2022 as the generational cutoff: before that date, most AI capabilities belonged to the previous generation; today’s tools include Doubao, DeepSeek, Yuanbao, Kimi, Emax, ChatGPT, Gymnast, Grok and Claude.
The qualifier 庄明浩 repeatedly emphasized was “looks.” This generation of AI resembles the Monkey King who wreaked havoc in heaven and defeated 100,000 celestial soldiers, but power does not equal reliability. Students are not meant to study how to build Sun Wukong; they are meant to learn how to be Tang Sanzang, the one who gives it direction.

7. Data, Algorithms and Compute Forge AI at the Cost of Massive Resource Consumption

To help students understand training, 庄明浩 compared text, images, video, voice and other records from the physical world to “massive amounts of data” fed into a furnace. Text allows knowledge to accumulate over time, and much of the knowledge students encounter today is built on written material.
He moved from a phone photo of roughly 2-3 MB to GB, TB and PB, then used Microsoft Flight Simulator’s simulation of airports, terrain and weather around the world as an example, saying its data volume reaches 2.5 PB. The point was to turn “massive” from an abstract adjective into an order of magnitude students could convert step by step.
Data alone does nothing. Algorithms digest the sprawling information, while compute supplies the energy needed for that digestion. He therefore had the class repeatedly memorize the “recipe for the Three-Flavored True Fire”: “data, algorithms, compute.”
By comparison, the human brain consumes roughly 20 watts yet can learn, transfer knowledge and generalize; current AI’s electricity consumption is “already enormous.” 庄明浩 did not equate machines with the human brain, but used the efficiency gap to show that similar-looking outputs may come from entirely different implementation paths.

8. AI Is a Probabilistic Parrot, Making Precise Inputs More Important Than Blind Trust

庄明浩 summarized large models as “a super-parrot that has read every book in the world”: they imitate and predict rather than think with feelings. When they see “The weather today is really—,” most people will fill in “good”; AI likewise guesses, character by character, the continuation that most resembles human language based on the distribution it learned.
When the prompt was reduced to “The weather today,” the students’ continuations immediately diverged. The classroom experiment established the key causal relationship: the vaguer the question, the more dispersed the candidate answers, and the more likely they are to be unsuitable for the person asking. The more precise the prompt, the less room the model has to guess.
Probabilistic generation produces “hallucinations”—in other words, “confidently talking nonsense.” AI has no feelings, morality or values; if the input is biased, the output may follow the bias. It does not truly understand that one plus one equals two, but has seen that combination countless times, and can still make common-sense errors that look absurd to humans.

9. “Role, Task, Requirements” Is a Transferable Prompting Formula

The worst prompt is “help me write an essay”: it provides no topic, audience, length or purpose, so AI can generate something only by guessing across a wide space. A better version asks it to act as an elementary-school Chinese teacher, produce 3 essay outlines on “an unforgettable weekend,” and make them vivid, full of rises and falls, and suitable for sixth graders.
The 3 elements serve different functions: role specifies whose perspective and capabilities to use; task says whether to explain, revise or provide ideas; requirements further constrain language difficulty, format, word count, quantity, tone, and whether to “use analogies” or “do not use analogies.”
庄明浩 believes most people think of role and task after using AI for a while; the genuinely difficult part is stating the requirements clearly. Ask it to explain a black hole with no further guidance, and it will produce a string of awkward technical terms. Require “an astronomer explaining black holes to sixth graders in plain language and with analogies,” and the result will be completely different.

10. A Good Learning Assistant Provides Scaffolding Without Doing the Thinking for the Student

庄明浩 defined “asking directly for the answer” as an absolute red line: receiving the answer immediately skips the student’s own thinking practice. When stuck on a math problem, students should ask AI to “act like a teacher and ask questions step by step to guide my thinking” until they find the method themselves, rather than requesting the final number.
For writing, students can ask AI for 5 science-fiction ideas about “schools of the future,” while explicitly saying “do not write the body.” If “happy” sounds too plain, they can ask for 5 idioms or elegant phrases. AI handles brainstorming and expanding the available options; students still choose and organize the work.
English corrections can be requested in table form, with errors on the left and the correct wording plus Chinese explanation on the right. For speaking practice, AI can play a restaurant server while the student plays the customer, with the instruction “say only 1 sentence at a time, and wait for my answer before continuing,” turning infinite generation into a manageable practice rhythm.
For concepts they do not understand, students can keep lowering the bar: “explain it in a way a sixth grader can understand,” then “in a way a 6-year-old can understand,” or ask for the Pythagorean theorem to be explained with LEGO. 庄明浩’s conclusion was: “AI will always be your apprentice and your tool.” 小庄 added: “Its power depends on your wisdom; only when you become stronger can it become stronger.”

11. AI First Compresses Execution Time, While Career Replacement Still Has No Certain Answer

庄明浩 used a sketch of a winged refrigerator to show the change. In the past, finding a professional illustrator or design agency might have taken 3 days to produce an advertising image; every revision to the color, font or wing size would add more time. Now, the same task may take only a few seconds, with AI accepting continuous revisions without emotion.
The question that follows is whether the people responsible for execution in the middle will disappear, and whether learning to draw will still matter. He used the expensive aristocratic portraits produced before the camera as an example: photography eliminated some highly paid portrait-painting work, but he did not force the analogy into a conclusion, saying plainly, “There is no answer to this question.”
His actionable message to students was: “AI will not replace you, but students who know how to use AI will find it easier to outperform those who do not.” But when asked whether AI will ultimately replace humans, he again preserved uncertainty: both camps—those who say it “definitely will” and those who say it “never will”—are strong. By the time the true “AI generation” grows up, no one knows the answer today.

12. Doubao Phone Turns Natural Language from a Question-and-Answer Interface into an Execution Interface

小庄’s first thought was spending less time taking out and staring at a phone: while driving or doing something else, users could simply state a task directly and avoid distraction. He believes this capability must understand complex requests better than Siri; “if Doubao Phone could get authorization for every app, I think it would be better.”
At the recording of “No Kidding,” 庄明浩 instructed it to open Xiaoyuzhou, search for “No Kidding,” find the first episode and like it. The show already had 200-300 episodes, and the phone advanced only about 3 episodes per screen, so everyone set it aside and continued chatting, agreeing to “keep recording until it reaches the first episode.” It eventually did complete the like.
Users can watch it click through each step, or let it continue in the background, with the screen off or through Dynamic Island. The demo also exposed the current practical boundary: the task can be completed, but scrolling is very slow. Whether it can execute more in-app tasks depends on the scope of its authorization.
庄明浩 compared it with the childhood science-fiction series Knight Rider. The intelligent car that once seemed impossible has now been partly realized through voice control of air conditioning, seats, routes and restaurant recommendations. Technology is no longer just answering questions; once granted permission, it can sense the outside environment, control devices and complete tasks autonomously.

A.I. Scientists Are Here. But Is Progress Accelerating? | EP 170 Gemini 3’s Comeback, Agent Models, and the RL Startup Opportunity